Semantic gain graphs improve multi-hop evidence retrieval for question answering
Retrieving Complementary Evidence for Multi-Hop RAG with Semantic Gain Graphs
Information Retrieval
Summary
Finding useful information spread across multiple documents is hard because some documents add new facts while others are just repetitive. The authors created a system called SG²RAG that uses a large language model to measure how much new information one piece of text adds when combined with another. They organize these measures into a graph and then use a special ranking method to find the most helpful documents for answering questions. Their method works better than others on several question-answering tests.
What this means in practice
- •For search engine developers: Enhance multi-document retrieval to find complementary evidence for complex queries, improving answer accuracy beyond single relevant documents.
- •For enterprise knowledge management teams: Improve automated retrieval of supporting documents that add unique insights when combined, aiding comprehensive report generation.
Authors
Zeliang Li, Xiaofen Xing, Kailing Guo, Xiangmin Xu
Abstract
Multi-hop retrieval-augmented generation (RAG) requires finding documents whose evidence is useful in combination. Query relevance and graph connectivity help locate related content, but neither directly measures the additional information one document contributes given another. We introduce SG$^2$RAG, a framework that represents this complementarity through semantic gain. Offline, a large language model assesses the relevant and nonredundant information contributed by one chunk given another within neighborhoods proposed by semantic similarity. These ordered assessments are aggregated into symmetric edge weights to construct a reusable Semantic Gain Graph. Online, controlled Personalized PageRank starts from query-relevant seeds and ranks complementary evidence beyond the initial retrieval set. Experiments on MuSiQue, 2Wiki, and HotpotQA with two language model backbones show that SG$^2$RAG achieves the best average EM, F1, and Recall@10 among the evaluated single-retrieval methods. Under identical candidate support and propagation, semantic gain improves average F1 and Recall@10 over similarity weighting by 4.08 and 4.31 points. Analyses of supporting documents and improvements in three iterative retrieval frameworks further support its effectiveness as a reusable retrieval component.